When the Map Starts Giving Orders

Joshua Lee returned a rented Tesla Model 3 to Hertz and received a $277.39 fuel-purchase charge. The electric car had no gasoline tank. According to the final receipt reviewed by automotive publication The Drive, its battery was at 96 percent when he collected it and 96 percent when he returned it. Hertz’s separate EV policy allowed cars to be returned at their starting charge level without a recharging fee. 1

Lee queried the bill. In a customer-service email published in May 2024, Hertz declined to adjust it because “the service was provided and contract is closed.” The response pointed to the fuel option in his signed rental agreement. 1

The mistake is funny until you reach the reply. An incorrect charge is easy enough to imagine. Having to dispute it with someone who cites the completed transaction as grounds for refusing a refund is the familiar, exhausting part.

The reporting does not establish that AI handled this complaint. But consider the task facing a team introducing an AI assistant into a service like this. Finding the charge, retrieving the agreement and explaining the policy could all produce a confident answer. The customer is asking whether the charge belongs there in the first place. That requires the assistant to examine something the organisation has already accepted.

A rental company needs records and standard procedures to operate. Staff cannot reconstruct every transaction from scratch whenever someone asks a question. The difficult design decision is when to reopen a question that the normal process has marked as settled. To understand why that can be harder than it sounds, it helps to look at an earlier effort to make a complicated business manageable.

A forest you could put in a ledger Link to heading

In eighteenth-century Prussia and Saxony, administrators needed to estimate how much timber their forests could supply and what revenue it would bring. Trees differed in species, age and size. Planting uniform stands made counting, forecasting and harvesting easier. 2

In Seeing Like a State, James C. Scott follows the consequences. The people calculating timber revenue also decided what to plant and clear. Other plants, local uses and ecological relationships barely registered in the account. In his telling, early commercial success gave way to declining forest health and productivity as those omissions caught up with the scheme. 2

Foresters themselves challenged the simplifications. Historian Jimmy Jönsson describes Henrik Hesselman’s investigations of soils and mixed forests in early twentieth-century Sweden. Hesselman wanted lasting yields; he argued that achieving them required closer attention to local ecological conditions. 3

That meant revising methods people had learnt, trusted and built careers around. A profession could agree on the importance of timber production and still disagree fiercely about what competent forestry required. The methods that solved an earlier problem now had to answer for consequences their users had learnt to recognise later.

The rules we inherit Link to heading

There is a question of timing here. The forester who remembers unreliable harvest estimates and the successor confronting damaged soil have encountered different reasons to worry. Each can struggle to understand why the other’s concern deserves priority.

Generations, William Strauss and Neil Howe’s 1991 account of American history, puts this relationship between experience and historical change at its centre. They propose four recurring generational types, shaped by the conditions people encounter as they grow up. As those people become parents and leaders, they help create the conditions that shape later generations. 4

In their model, people shaped by a crisis help build institutions around collective action and stability. A later awakening challenges established values and authority; fragmentation eventually gives way to renewed demands for collective action. These changes help produce the conditions for the cycle to recur. 4

For digital design, a useful extension of this idea is to examine how a successful reform changes the circumstances in which its rules will operate. People who built an institution to prevent a familiar disaster may value its consistency and authority. Their successors inherit its protections, but also encounter its restrictions, exclusions and failures. They can have good reasons to ask for change without sharing the experience that made the original arrangement seem indispensable.

Any reform they achieve starts the process again. It creates procedures, rewards certain behaviour and leaves somebody responsible for keeping it running.

That does not make every challenge wise. A safeguard can work so well that people forget what it prevents. Nor do birth dates settle these disagreements; Pew Research Center’s caution about generational labels stresses the differences within age groups and the difficulty of separating cohort effects from age and shared events. 5 The useful move is to investigate what made a practice sensible, what it still protects and what has changed.

Software makes that investigation easy to postpone. An organisation can replace an interface while retaining the same categories, permissions and definition of a completed case. The reasons behind those decisions may be poorly remembered, but the workflow still enforces them. An AI assistant introduced into that workflow has an institutional inheritance before it answers its first customer.

What an AI assistant inherits Link to heading

That inheritance can include policy manuals, examples of past responses and records of cases labelled successfully resolved. A team using these materials to configure and evaluate an assistant needs to ask what “successful” meant. Did the customer receive a correct outcome, or did the file simply stop requiring attention?

Hertz later attributed the charges to a systems error without identifying its precise cause. 6 We cannot trace it to an outdated rule or an AI deployment. What the correspondence does show is a response that treated the completed transaction and its terms as sufficient grounds to refuse an adjustment.

Suppose a team automating a similar service asks its assistant to explain charges using billing records and rental terms. It may produce an accurate account of company policy while never checking whether the disputed entry belongs on the bill. If the evaluation rewards finding the right policy and closing the enquiry, the team has built that omission into its definition of good work.

The same technology could support a broader investigation. With appropriate access, an assistant could compare the vehicle record, charge and terms, then identify a contradiction for someone authorised to act. But those records make different kinds of claims. Bringing them together only helps if the investigation distinguishes what each can establish.

Several accounts, one transaction Link to heading

Three records bear on the charge, and each answers a different question.

Each contributes something to the investigation, but they do not answer the same question.

An agreement cannot, on its own, establish that the described service occurred. Nor does a customer’s objection automatically prove an error. The investigator has to connect the terms to the actual transaction and check the evidence supporting the charge.

That sounds straightforward when the disputed service is gasoline for an electric car. It gets harder when a fee has an ambiguous label, a record is incomplete or several policies could apply. A useful assistant can help identify what would settle the disagreement: a missing record, a mistaken entry, an exception in the terms. Inventing plausible reasons for the fee would add another account requiring verification.

Looking through several lenses therefore requires taking a position on the evidence. If the vehicle record contradicts the description of the service, the assistant should bring that conflict into the investigation. Producing a sympathetic summary of both sides would leave the customer exactly where they started.

The company can issue the bill and control how it is corrected. Unless contrary evidence can change that process, a conversational interface gives customers more room to explain themselves while leaving the company’s answer effectively settled. And even correcting this customer’s bill leaves a further question: what happens to the next person with the same problem?

When the exceptions keep coming Link to heading

In Lee’s case, The Drive published the dispute, and Hertz then said it would apologize and refund the charge. Further reporting found other EV renters who had received refueling charges. Hertz acknowledged a systems error, said it had fixed the problem and reported refunding affected customers. 6

A company designing automated support should ask how evidence of a recurring problem reaches that second level before a customer needs a journalist.

An assistant could help connect complaints that otherwise sit in separate files, preserve the evidence behind each and flag a pattern for investigation. Repetition would justify a closer look. It would not establish the cause: a sound rule may be implemented badly, information may be missing, or the rule itself may no longer fit the transactions it governs. Someone has to distinguish those possibilities before deciding what to change.

This is where friction belongs: inside the organisation. Repeated contradictions should create work for the people responsible for the process. They need to examine why it exists, which protections remain necessary and whether its definition of success is still adequate. Closing each complaint separately can conceal that work, especially if the dashboard treats a refund as another completed case.

The connection to Generations becomes useful here. People inherit arrangements built in response to earlier experience; preserving them and replacing them both have consequences. A product team should be able to explain what a rule still achieves and what evidence would warrant revising it. An AI system that helps assemble that evidence could make reconsideration easier. An evaluation that counts only compliance with existing procedures would miss that contribution.

For a support team, a practical place to start is the queue of cases resolved through exceptions.

1. Put the underlying complaints together.
2. Check whether the same condition keeps forcing staff to work around the process, and whether people who never complained were affected too.
3. Give those findings to someone who can change the rule, the software or the measure of success.
4. Check whether the next comparable case still requires the customer to fight for a correction.